From Flat Facts to Sharp Hallucinations: Detecting Stubborn Errors via Gradient Sensitivity

Researchers propose Embedding-Perturbed Gradient Sensitivity (EPGS), a detection method that identifies "stubborn hallucinations"—high-confidence factual errors in LLMs—by measuring gradient magnitude spikes after embedding perturbation, exploiting the hypothesis that robust facts occupy flat loss minima while brittle memorization occupies sharp minima.